Examples

You can find examples here.

Basic examples

• Quick start.

• MNIST example.

• MNIST VAE example.

• CIFAR example.

• Model loading.

• Sequential LR scheduler.

Advanced examples

• CIFAR with DPP, mixed precision and gradient accumulation.

  Single GPU training:

  python cifar_advanced.py --batch_size 256 --lr 0.001
  

  Single machine 2 GPUs distributed data parallel training:

  ./cifar_advanced.sh 2 --batch_size 128 --lr 0.0005
  

  DDP training with mixed precision and gradient accumulation:

  ./cifar_advanced.sh 2 --batch_size 128 --lr 0.0005 --amp --iter_size 2
  

• Custom callback events.

• Custom build methods for creation of model parts.

Kaggle solutions

• 1st place solution for Freesound Audio Tagging 2019 (mel-spectrograms, mixed precision)

• 14th place solution for TGS Salt Identification Challenge (segmentation, MeanTeacher)

• 22nd place solution for RANZCR CLiP - Catheter and Line Position Challenge (DDP, EMA, mixed precision, pseudo labels)

• 50th place solution for Quick, Draw! Doodle Recognition Challenge (gradient accumulation, training on 50M images)

• 66th place solution for Kaggle Airbus Ship Detection Challenge (instance segmentation)

• Solution for Humpback Whale Identification (metric learning: arcface, center loss)

• Solution for VSB Power Line Fault Detection (1d conv)

• Solution for Bengali.AI Handwritten Grapheme Classification (EMA, mixed precision, CutMix)

• Solution for ALASKA2 Image Steganalysis competition (DDP, EMA, mixed precision, BitMix)

previous

Guides

next

API reference